Skip to content

scalpels-app/scalpels

v0.3.0MIT

One connection to your self-hosted Scalpels apps, plus skills for running them well.

Scalpels for Claude and Codex

One connection to your Scalpels apps, plus skills that teach your agent how to run them well.

Scalpels is a catalog of free, MIT-licensed Laravel apps that you deploy to your own Laravel Cloud or Forge account. This plugin connects your agent to the Scalpels MCP server, which can read the catalog and docs, plan and run deployments, connect existing sites, and relay questions to the apps your team has installed.

Install

Claude Code

claude plugin marketplace add scalpels-app/scalpels-plugin
claude plugin install scalpels@scalpels

Codex

With a recent Codex CLI that supports codex plugin:

codex plugin marketplace add scalpels-app/scalpels-plugin
codex plugin add scalpels@scalpels

For a local checkout, use codex plugin marketplace add /absolute/path/to/scalpels-plugin instead of the GitHub source.

GitHub Copilot CLI and other Agent Plugins hosts

This repository is also an Agent Plugins 1.0 package, so hosts that support that format can install it from the repository. In GitHub Copilot CLI:

copilot plugin marketplace add scalpels-app/scalpels-plugin
copilot plugin install scalpels@scalpels

Every client uses the same skills/ directory and the same MCP endpoint; installing in one client does not install in the others. Do not add the MCP server separately as well.

Authorize the Scalpels OAuth connection when prompted by your client so you can sign in and approve access. There are no tokens to copy.

For headless or CLI setups, create a token on the Agent access page in Scalpels instead. See the agent docs.

What's included

ComponentWhat it does
scalpels MCP serverhttps://scalpels.app/mcp, authenticated with OAuth
scalpels-fleet skillOrients your agent in your teams and installed apps, and sets the ground rules for deploying, connecting sites and calling apps
hone-investigate skillChecks Hone coverage, errors, slow routes and abnormal telemetry per app; proposes fixes without making changes
Agent Plugins, Claude and Codex manifests and marketplacesInstall the same plugin from this repository in any of these clients

What your agent can do with it

The OAuth connection grants the mcp:use scope. With it, your agent can:

  • read your teams, sites, deployments, the catalog and docs
  • connect a Forge site or Cloud environment to a product you run, which writes that product's keys into the site's environment
  • plan and start deployments to your Laravel Cloud account
  • call your installed apps' tools as you, with your role in each team

Each team still enforces your current role. Revoking the connection on the Agent access page stops the current token and its ability to refresh.

Unattended jobs

A scheduled job that only collects telemetry or investigates problems doesn't need to deploy, connect sites or change anything in your apps. MCP clients hold one connection per server, so the connection itself can't be narrowed per job. Use your client's per-tool permissions instead:

  • Allow the read tools (app_read, installed_apps, app_tools, list_deployments, deployment_status, deployment_drift, docs) without asking.
  • Set app_write, app_call, deploy_app, connect_site, resume_deployment, isolate_app and restore_app to ask.
  • Set app_confirm (destructive app actions) to ask or never.
  • Run scheduled tasks without automatic approval. A job that reaches a write then stops at the approval prompt instead of acting on its own.

This works because Scalpels routes every app call through one of three doors (app_read, app_write, app_confirm) according to what the tool does, so the client can gate each door separately.

Where your data goes

Your apps and their data live in your own Laravel Cloud or Forge account. Scalpels holds no product data.

When your agent calls one of your installed apps, the request and response pass through Scalpels' memory on the way. They are not end-to-end encrypted. Scalpels does not intentionally store that content; it records who asked which app for what, when, and whether it succeeded.

If you'd rather not route through Scalpels at all, point your client directly at your apps' own MCP endpoints. Fork button is up and to the right.

Privacy · Terms — both policies, published October 30, 2026, are live on scalpels.app and are configured in both manifests.

Package references

The root plugin.json and mcp.json follow the Agent Plugins 1.0 specification. Codex, ChatGPT and other OpenAI surfaces read the portable package; OpenAI's listing metadata lives under extensions.com.openai in plugin.json, as OpenAI's package guide and manifest field reference describe. The Codex marketplace at .agents/plugins/marketplace.json uses the documented marketplace format.

Claude Code doesn't read the Agent Plugins files. It uses .claude-plugin/plugin.json, .claude-plugin/marketplace.json and .mcp.json. .mcp.json and mcp.json declare the same endpoint; Claude needs "type": "http" and Agent Plugins needs "type": "streamable-http", so keep both URLs in step. OAuth endpoints, dynamic client registration and the mcp:use scope are discovered from the server's OAuth metadata. No client secrets or bearer tokens are packaged.

The listing icon is the existing Scalpels mark from public/favicon.svg in artisan-build/scalpels.app, rasterized without redrawing to a 1024 × 1024 PNG with ImageMagick:

magick -background none -density 192 ../scalpels.app/public/favicon.svg -resize 1024x1024 -strip .claude-plugin/icon.png

Review the icon before the plugin is first saved or submitted in Claude's developer portal: that first save sets the listing icon, and later changes to the file don't update it. A successful local validation does not replace the directory's safety scan or policy review. Claude's directory can validate and review a private repository, but the repository must be public before the listing goes live; that visibility change and all directory submissions remain human steps.

License

MIT